On July 28, BYD held a closed-door media roundtable attended by outlets including 36Kr. The session featured Yang Dongsheng, head of BYD’s new technology institute, deputy director Han Bing, and several technical leads.
When asked about the impact of BYD’s recent smart driving initiatives, Yang told 36Kr that the company has observed “a clear rise in driver confidence” since launching the parking safety guarantee.
He noted that assisted parking usage has increased from 30–40% to over 60%. While adoption of city navigation-on-autopilot (NOA) remains lower, at 20–30%, due to lingering trust issues, highway NOA usage has reached peaks of 90%.
In February, BYD introduced 21 new car models equipped with advanced driver assistance systems (ADAS), built on three DiPilot platforms: 600, 300, and 100.
- DiPilot 600 targets BYD’s premium Yangwang brand and features dual Orin X chips delivering 508 TOPS (trillions of operations per second) of computing power, along with three LiDAR (light detection and ranging) sensors.
- DiPilot 300, used by the Denza and BYD marques, incorporates a single Orin X chip and up to two LiDAR sensors.
- DiPilot 100, BYD’s most broadly deployed solution, omits LiDAR entirely. It combines Nvidia’s Orin N platform with Horizon Robotics’ J6M chip and is positioned for cost-effective scaling.
Current capabilities include highway and expressway navigation and assisted parking. BYD plans to introduce city memory-based navigation by year’s end.
Smart driving capabilities now span BYD’s full pricing spectrum, from the Song L to the Seagull.
In early July, all 21 assisted driving-equipped models received a major update. A key highlight was BYD’s pledge to cover all repair and service costs for accidents occurring during automated parking, insulating users from insurance liabilities tied to such incidents.
Despite the rapid rollout, BYD asserts its core strength remains in its powertrain and chassis systems, which underpin its assisted driving features.
The DiSus-A and DiSus-C systems are already in mass production. Yang said these reduce parking time by 50–80%. He also referenced a recent demonstration of improved maneuverability in tight urban U-turns, though broader deployment will require further validation.
BYD remains conservative on vision-language-action (VLA) models. Li Feng, head of intelligent software development, said VLA is not yet mature and has not outperformed existing end-to-end or vision-language model (VLM) architectures. Accordingly, BYD is not heavily investing in VLA at this time.
As for custom chips, Yang noted BYD currently has no roadmap for self-developed high-performance computing. “The real challenge in assisted driving is algorithms and data, not just compute power,” he said. “Our focus is on algorithms that run effectively on midrange chips at 100, 300, or 500 TOPS.”
While BYD continues to develop its proprietary assisted driving stack, it remains open to partnerships. “Self-reliance deepens our technical understanding,” Yang said, “but we’re large enough to benefit from both in-house development and external collaboration. This two-track approach will continue.”
The company maintains that long-term adoption hinges more on safety and usability than on aggressive feature expansion. Yang emphasized that trust grows through reliability, especially in mass market segments. “Our aim is to integrate assisted driving into daily routines, even for owners of vehicles priced below RMB 100,000 (USD 14,000).”
The following transcript has been edited and consolidated for brevity and clarity.
Q: What impact has assisted driving adoption had?
Yang Dongsheng (YD): Parking usage has risen to over 60%, up from 30–40%. Highway NOA reaches 90% at peak. City NOA varies based on user trust and traffic complexity but has grown from 10% to around 20–30%.
Q: Has DiPilot 100 influenced sales?
YD: It’s still early. It took a decade to mainstream electrification. Smart driving will spread faster but needs time, especially for buyers under RMB 100,000. Our parking guarantee policy has improved trust and usage behavior.
Q: What is BYD’s stance on VLA models and custom chips?
Li Feng: VLA hasn’t outperformed current methods, so we’re not prioritizing it.
YD: Larger compute is helpful, but we’re focused on optimizing models for midrange chips. That’s more scalable. We have no current plans to develop custom chips.
Q: How do you balance in-house and external development?
YD: We have the scale for both. Internal work gives us control; partnerships offer agility. Both are important for long-term success.
Q: What’s your approach to training data?
YD: We focus on identifying “expert drivers.” If someone drives consistently well, about 80% of their trips become valuable training data. We’ve built tools to detect these patterns and are integrating artificial intelligence for data labeling.
Q: How do BYD’s chassis and assisted driving features complement each other?
YD: Our strength lies in chassis and powertrain design. DiSus-A and DiSus-C features are more than visual gimmicks. They reduce parking time by up to 80%. We’ve recently shown this in tight U-turn tests. Broader deployment is pending further validation.
Q: How is BYD addressing concerns over assisted driving safety?
YD: Parking and highway driving capabilities have become relatively mature. Urban driving in China is still chaotic as electric scooters, unpredictable signals, and road behavior are major variables. Software alone can’t solve this. Tire grip remains critical. In the combustion-based car era, slip was hard to measure. Electric motor control lets us measure and adjust output precisely, improving safety. Smart driving today is about merging electrification with intelligence, and that’s our current focus.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Xu Caiyu for 36Kr.